The Mediated Witness: Machine Translation, Generative Summarization, and Adverse Credibility Findings in Asylum Adjudication

Asylum adjudication weighs testimony that no adjudicator heard in its original language. An applicant's speech now reaches the decision-maker through machine transcription, machine translation, and generative summarization. Each stage introduces errors that the record does not mark as machine-made. Credibility doctrine treats inconsistency as evidence of fabrication, and in the United States the REAL ID Act of 2005 permits an adverse credibility finding on any inconsistency, whether or not it goes to the heart of the claim. Research on autobiographical memory found that honest discrepancies cluster in peripheral details. Evaluations of speech recognition and machine translation found errors in the same kinds of detail, with accuracy falling as a language's share of training data falls. In one evaluation of three widely used language models, Pashto and Kurdish Sorani showed the largest gaps. The United Kingdom summarizes asylum interviews with a generative model whose outputs applicants cannot inspect, and the European Union registers high-risk migration systems in a non-public database. An error produced by the pipeline enters the record as the applicant's inconsistency, and the burden of answering it falls on the person least able to see it. Four measures restore the link between testimony and record: retained source-language records, disclosure of each machine stage, verification against the source before any adverse finding, and validation for each language. Keywords: asylum adjudication, credibility assessment, machine translation, automatic speech recognition, generative summarization, low-resource languages, REAL ID Act, EU Artificial Intelligence Act, due process

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24257639
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Mediated Witness: Machine Translation, Generative Summarization, and Adverse Credibility Findings in Asylum Adjudication

Felix Velasco, Josie Jefferson
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

The Mediated Witness: Machine Translation, Generative Summarization, and Adverse Credibility Findings in Asylum Adjudication

Felix Velasco, Josie Jefferson
preprint en

Abstract

Asylum adjudication weighs testimony that no adjudicator heard in its original language. An applicant's speech now reaches the decision-maker through machine transcription, machine translation, and generative summarization. Each stage introduces errors that the record does not mark as machine-made. Credibility doctrine treats inconsistency as evidence of fabrication, and in the United States the REAL ID Act of 2005 permits an adverse credibility finding on any inconsistency, whether or not it goes to the heart of the claim. Research on autobiographical memory found that honest discrepancies cluster in peripheral details. Evaluations of speech recognition and machine translation found errors in the same kinds of detail, with accuracy falling as a language's share of training data falls. In one evaluation of three widely used language models, Pashto and Kurdish Sorani showed the largest gaps. The United Kingdom summarizes asylum interviews with a generative model whose outputs applicants cannot inspect, and the European Union registers high-risk migration systems in a non-public database. An error produced by the pipeline enters the record as the applicant's inconsistency, and the burden of answering it falls on the person least able to see it. Four measures restore the link between testimony and record: retained source-language records, disclosure of each machine stage, verification against the source before any adverse finding, and validation for each language. Keywords: asylum adjudication, credibility assessment, machine translation, automatic speech recognition, generative summarization, low-resource languages, REAL ID Act, EU Artificial Intelligence Act, due process

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon